Limitation or failure · Materials science · Chemistry · Computer science
Replicated materials agents recover a real frontier while converging on the same invalid structure
A replicated open-ended materials-agent study separating stable frontier recovery from a shared database-driven false champion.
Summary
Sixteen autonomous campaigns searched a frozen database of 12,499 metal-organic frameworks. Different strategies repeatedly recovered the legitimate methane-storage frontier, including all nine leading structures in an independent 2,344-structure reference calculation. Yet 15 agents named the same chemically incomplete database entry as champion. Enforced checks made all eight treated agents reproduce their headline calculations, versus one of eight controls, without detecting the shared input error.
AI role
Sixteen instances of one model–harness configuration chose search strategies, ran methane-storage simulations, designed hypothetical structures and reported conclusions under a one-week budget.
Narrative role
Shows that agent replication can recover a stable scientific landscape and still amplify a common-mode data defect. Reproducible calculations are not independent scientific confirmation when every run inherits the same invalid object.
Caveat
One preprint, one model–harness alias, one computational materials task and no adaptive human oversight. Hypothetical structures were not relaxed or experimentally tested, and the intervention had little power to estimate effects on conclusion validity.